AgeFlow: Conditional Age Progression and Regression with Normalizing Flows

Zhizhong Huang, Shouzhen Chen, Junping Zhang, Hongming Shan

Introduction

Synthesizing the facial aging process—both progression (aging) and regression (rejuvenation)—is a long-standing challenge in computer vision. While Generative Adversarial Networks (GANs) have achieved significant milestones, they often struggle with training instability, attribute drift (e.g., unintended changes in gender or ethnicity), and a lack of mathematical reversibility. AgeFlow addresses these issues by combining the stability of Normalizing Flows with the generative power of GANs. By employing an entirely invertible architecture, AgeFlow ensures a bijective (one-to-one) mapping between age groups, effectively preserving identity and facial attributes.

Methodology: The AgeFlow Architecture

The core of AgeFlow is its focus on bijectivity. Unlike traditional GANs, which map from a low-dimensional noise space to a high-dimensional image space, AgeFlow utilizes a flow-based generator to create a perfectly reversible path between the image and its latent representation.

The AgeFlow architecture features an invertible encoder/decoder and an Invertible Conditional Translation Module (ICTM) to ensure bijective age transformations.

The framework consists of three primary components:

  1. Invertible Encoder/Decoder: Based on a flow-based generator ($G$), this component maps input images into a latent space. Its mathematical inverse ($G^{-1}$) serves as the decoder, ensuring no information is lost during the encoding process.
  2. Invertible Conditional Translation Module (ICTM): This module performs the actual age transformation within the latent space. Because the ICTM is also invertible, the transformation between a "young" latent vector and an "old" one is one-to-one.
  3. Attribute-aware Knowledge Distillation: To maintain consistency in non-age attributes like race and gender, the model uses "average aging directions" from a pre-trained teacher model to guide the ICTM.
  4. Latent Space GAN: Training stability is enhanced by placing the GAN discriminator in the latent space rather than on raw pixels, ensuring that translated latent vectors remain indistinguishable from real vectors of the target age group.

Experimental Evaluation

AgeFlow was trained and evaluated on the MORPH and CACD (Cross-Age Celebrity Dataset) benchmarks. To test generalization, the authors also performed external testing on the FG-NET and CelebA datasets, demonstrating the model's effectiveness on "in-the-wild" images.

Key Results

The model achieved state-of-the-art performance across several key metrics:

  • Age Accuracy: On the MORPH dataset, AgeFlow reached over 90% accuracy in target age classification, outperforming previous models like IPCGAN.
  • Identity Preservation: Using ArcFace cosine similarity, the model maintained high identity consistency (scores of 0.8-0.9), a significant improvement over earlier GAN methods which often dropped to 0.2-0.6.
  • Visual Quality: The model produces realistic aging effects, such as wrinkles and graying hair, while keeping the underlying facial structure intact.
Qualitative comparison of age progression and regression, demonstrating high-fidelity aging effects and identity preservation.

Conclusion and Future Work

AgeFlow demonstrates that bijectivity is a powerful tool for identity-preserving image translation. By operating in a structured latent space provided by Normalizing Flows, the model avoids the common pitfalls of GAN-based aging. While the model currently focuses on age, gender, and race, future work may involve modeling more granular attributes like hairstyle or expressions. Additionally, the "Flow + Latent GAN" approach holds promise for other image-to-image translation tasks beyond the domain of facial aging.

Summary updated

Abstract

Age progression and regression aim to synthesize photorealistic appearance of a given face image with aging and rejuvenation effects, respectively. Existing generative adversarial networks (GANs) based methods suffer from the following three major issues: 1) unstable training introducing strong ghost artifacts in the generated faces, 2) unpaired training leading to unexpected changes in facial attributes such as genders and races, and 3) non-bijective age mappings increasing the uncertainty in the face transformation. To overcome these issues, this paper proposes a novel framework, termed AgeFlow, to integrate the advantages of both flow-based models and GANs. The proposed AgeFlow contains three parts: an encoder that maps a given face to a latent space through an invertible neural network, a novel invertible conditional translation module (ICTM) that translates the source latent vector to target one, and a decoder that reconstructs the generated face from the target latent vector using the same encoder network; all parts are invertible achieving bijective age mappings. The novelties of ICTM are two-fold. First, we propose an attribute-aware knowledge distillation to learn the manipulation direction of age progression while keeping other unrelated attributes unchanged, alleviating unexpected changes in facial attributes. Second, we propose to use GANs in the latent space to ensure the learned latent vector indistinguishable from the real ones, which is much easier than traditional use of GANs in the image domain. Experimental results demonstrate superior performance over existing GANs-based methods on two benchmarked datasets. The source code is available at https://github.com/Hzzone/AgeFlow.
BibTeX
@inproceedings{huang2021ageflow,
  author = {Zhizhong Huang and Shouzhen Chen and Junping Zhang and Hongming Shan},
  title = {{AgeFlow: Conditional Age Progression and Regression with Normalizing Flows}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/103},
}
Figure